AI tools for marketing analytics: where numbers come from

A number in an AI analytics answer comes from one of three places: a query the tool ran, the model's own arithmetic, or an estimate. Only the first leaves something you can re-run. Google's help page says Ask Advisor in Google Analytics "may have some capability limitations and make mistakes", and Google's BigQuery documentation says Gemini for Google Cloud products can "generate output that seems plausible but is factually incorrect". Use AI to ask and to draft, and take the number that moves budget from a report, query or test you can repeat.

Which AI features already sit inside your analytics tools?

Five are documented by their vendors, as the pages read today:

AI does help where the output is a draft you can read: a query you then check, a report you already trust explained in plain words, finished results summarised. None of these tools changes where the number has to come from.

What do benchmarks and experiments show about AI on marketing questions?

Benchmarks score SQL and agents, not your export. AD-Bench (arXiv preprint by Tencent authors, revised 22 June 2026, not peer-reviewed) built 225 test requests from real marketing-analysis requests on a production advertising platform. The best model tested answered 76.9% correctly on the first try overall and 61.4% on the hardest level, graded by an LLM judge that was also one of the models tested, a limitation the authors list. On the Spider 2.0 leaderboard, as read today (a benchmark project's page, with entries submitted by named teams), the top Snowflake-track entry scores 96.70 on 547 examples that include "well-prepared database metadata and documentation". Nothing on that page says how those systems score on your own export.

"Why" questions come in two kinds. Google's page on AI overviews says its AI "works through countless combinations of dimensions and metrics to proactively connect the dots, explaining why these spikes happened", and Ask Advisor lists "Why did my total revenue drop on August 1st?" among its questions. Some of those can be checked against independent records: a tracking break, a stock-out, an outage. "Which channel caused these orders?" is the harder kind. Gordon and co-authors (Marketing Science, 2019, peer-reviewed) compared 15 US advertising experiments at Facebook, with 500 million user-experiment observations and 1.6 billion ad impressions, against observational estimates. They report that "observational methods often fail to produce the same effects as the randomized experiments". A tool that only sees your property's data can describe what changed. Until a holdout backs it, credit for a cause is a hypothesis.

How do you check an AI answer yourself?

Take one answer that matters and run five checks:

1. Sort each number by source. Ask the tool what produced it: a report, a query, or its own calculation. A number with no source is unverified.
2. Rebuild it. Open the same report with the same dates, time zone, dimension and filter, or run the SQL yourself and reconcile orders to Shopify for the same days.
3. Match the definitions. A rate per user and a rate per session are different numbers, and so are purchases and key events.
4. Ask again later. BigQuery's page says the same prompt can return different syntax, so repeat the question on another day and after an update.
5. Treat "why" as a hypothesis. Check it against independent records (a campaign change log, stock, outages) and, for channel credit, a holdout.

Pass: the same number, or a gap you can name. Fail: no source, an unexplained gap, or a different answer to the same question. Vendors grade their own assistants, so the accuracy figure that matters for your store is the one you get this way. Also ask where your data goes: Google says Ask Advisor data, "including prompts and generated answers, is not used for training, fine-tuning, or evaluation of AI models", and that chat logs are kept for 55 days. Ask any other tool for the same two facts before you paste order data into it.

Where does a causal read fit?

Later, once the question is which channel caused orders rather than what changed, a causal attribution tool like Causality Engine shows what each channel caused from a GA4 export, and its Pro plan includes AI chat across your historical attribution data.

Sources, 30 September 2026: Ask Advisor in Google Analytics (Beta) (Google Analytics Help, 2026); How Ask Advisor uses data (Google Analytics Help, 2026); AI overviews (Google Analytics Help, 2026); Google Analytics MCP server (Google, 2026); Write SQL queries with Gemini (Google Cloud, 2026); Sidekick (Shopify Help Center, 2026); Marketing reports (Shopify Help Center, 2026); AD-Bench (Hu et al., arXiv, 2026); AD-Bench full text (arXiv, 2026); Spider 2.0 leaderboard (XLANG Lab, read 30 September 2026); A Comparison of Approaches to Advertising Measurement (Gordon et al., Marketing Science, 2019).

Related answers

• Should you use an LLM to analyze your attribution data?
• Can ChatGPT replace your attribution platform?
• AI chat in attribution tools: what to ask the vendor
• Where LLMs actually help in marketing, with a check for each
• Incrementality testing, explained